Article

How to set up an AI agent: the first two weeks

Setting up an AI agent means telling it about yourself and your work, giving it simple first tasks and correcting it in words when it gets something wrong. All of this happens in an ordinary Telegram chat, without code. The agent writes what you say into memory, so every week it asks less and does better. Below is a plan for the first two weeks.

Updated October 14, 2026·7 min read
Neko on a balcony records a voice message, and the robot cat agent writes her words into a glowing notebook

What “setting up” an agent means

An agent is installed once: server, model, harness, Telegram. That is installation. Setup begins after it: the agent learns about you, your work and your rules. It is like a new employee getting up to speed.

You set up an agent with words. You type or dictate by voice, the agent writes the important parts into its knowledge base and relies on them in later tasks.

Day one: tell the agent about yourself

Start with an introduction. One long message or a few minutes of voice is enough. Cover five things.

  • Who you are and what you do

    What you work on and your roles: owner, accountant, parent. What a typical day looks like.

  • People

    Key clients, partners, staff and family: who is who and how to address them.

  • Rhythm

    When you work, when not to disturb you, what time suits you for summaries.

  • Boundaries

    What the agent does on its own and what only with your approval: client emails, payments, deleting files.

  • Style

    How to write to you: short or detailed, formal or casual, text or voice.

At the end, ask the agent to retell what it remembered. You will see right away what it got wrong, and that is easy to fix.

Week one: simple tasks

Pick first tasks where a mistake is easy to spot and costs nothing.

  • A summary: what is on today in your tasks and inbox, and what not to forget.
  • Reminders: “on Thursday, remind me to call the supplier”.
  • Drafts: an email to a client, a reply to a review, a post. You read and send them yourself.
  • Digests: send a document or a long email and ask for the gist in three points.

Every task is a chance to correct. The agent does not know your habits yet, and the first week goes into teaching them.

How to correct the agent

Correct it right away and specifically, as you would an employee. A vague correction the agent will interpret its own way; a specific one it will follow.

VagueClear to the agent
Tone“Write normally”“Write to clients formally, no emoji, five lines max”
Deadline“Remind me later”“Remind me on Friday at 10 am”
Result“Sort the inbox”“List client emails with no reply: who, about what, how many days they have waited”
Rule“Not like that”“Always send invoices to me for review, never send them yourself”

If a correction applies to all similar tasks, say so directly: “remember this as a rule”. The agent will write the rule into memory and keep applying it. A couple of days later, check: ask which rules it knows about you.

A robot cat climbs three steps up from a memory notebook: a conversation, a finished task, a schedule clock
Introduction, simple tasks, a schedule: the agent climbs the steps together with its memory.

Week two: scheduled tasks

Once the agent stops making mistakes on simple things, give it regular work. It does this on its own, without a message from you.

  • Every morning, a summary of the inbox and the day's tasks.
  • On Mondays, a to-do list for the week from your notes and agreements.
  • On Friday evening, which clients are waiting for a reply.
  • Once a month, a reminder about payments and renewals.

The schedule is also set in words: “every weekday at 9 am send me an inbox summary”. If the summary is inconvenient, ask for a different format.

How the value of memory grows

In the first days the agent often asks again: who Sergey is, what was agreed with a client, where to save a file. That is expected: the memory is still empty.

Every task adds people, decisions and results to the knowledge base. After a couple of weeks the agent recalls on its own what you agreed with a client last week and puts it into the email. In NekoAgent memory is an Obsidian knowledge base on your server: you can open and edit the notes by hand, and they survive restarts.

Common mistakes in the first two weeks

  • Giving everything at once. Ten tasks on day one, and it is unclear where the agent went wrong. Add them one or two at a time.
  • Silently redoing the agent's work. If you fixed an email and did not tell it, it will repeat the mistake.
  • Setting no boundaries. Keep client emails, payments and deleting files for your approval.
  • Giving up after three days. The value grows with memory, and the first days are the agent's weakest.

Setup with a turnkey install

You do not have to do the first setup alone. In NekoAgent the first install happens on a call with a curator, about two hours: together we connect the server, the engine and Telegram and show you how to work with the agent and how to train it further for your tasks.

After that, during paid support Iris monitors the agent around the clock and helps bring it back after failures (reasonable efforts, no guaranteed response time), makes nightly encrypted backups and ships tested updates you install with one button. Support is free for the first 30 days, then 750 RUB a month for every 5 agents.

How to install an agent yourself or turnkey is covered in How to create your own AI agent: DIY or turnkey, and how an agent differs from a chat in AI agent vs AI assistant.

Let's install and set up your agent together

Plans from 24,900 RUB for three installs. The first one is on a call with a curator, with training.

Key questions

How do I set up an AI agent?

Tell it about yourself, your people and your rules, give it simple tasks and correct it in words when it gets something wrong. After a week, add scheduled tasks. All of this happens in a chat, without code.

How long does setup take?

The introduction fits into one evening. After that, setup happens as you work: in the first two weeks you correct the agent more often, then corrections become rare.

Do I need to know how to code?

No. You set up the agent with ordinary Telegram messages, in text or voice. Only those who install the agent on a server themselves will need the terminal.

What should I do if the agent makes mistakes?

Say right away and specifically what is right, and ask it to remember this as a rule. Keep important actions on your approval until you are sure the agent does them correctly.

What should I tell the agent first?

What you do, who the key people are, when not to disturb you and what it may do without asking. It will learn the rest as it works.

More articles

Neko assembles an AI agent from parts: a model, a server, Telegram and a knowledge base
Article · 8 min

How to create your own AI agent: DIY or turnkey

Your own AI agent needs four things: a model, a harness that gives the model memory and tools, a server where the agent runs around the clock, and a channel to talk to it, usually Telegram. You can put this together in a no-code builder, by hand from open source, or turnkey. Here is what each path costs in money and time.

Read the article
  • What an agent is made of
  • Three paths compared
  • Monthly cost
  • Turnkey, step by step
Neko weighs free software on a scale against a server, a model and time
Article · 6 min

Free AI agents: what is really free and what you pay for

The only free part of an AI agent is the software: OpenClaw and Hermes Agent are open source. The server that keeps the agent running around the clock and the model that does its thinking always cost money, and you pay for installation and upkeep with your time. The free tier of ChatGPT or Claude gives you a chat assistant, and you cannot build an agent out of it.

Read the article
  • What is free and what is not
  • Free ChatGPT vs an agent
  • OpenClaw and Hermes
  • Monthly costs
Neko examines a transparent robot cat through a magnifying glass; inside are the model, tools, memory and schedule
Article · 6 min

How AI agents work: explained in plain words

An AI agent is a language model that has been given tools, memory and a schedule. The model decides what to do, and a harness program carries out its decisions: reads files, sends messages, runs commands on a server. The agent works in a loop: it gets a task, makes a plan, takes a step, checks the result and decides what comes next.

Read the article
  • What an agent is made of
  • The work loop
  • How an agent remembers
  • Where agents go wrong
Neko at a desk between two robots: the assistant holds a speech bubble, the robot cat agent heads off to work with letters, an alarm clock and a key
Article · 6 min

AI agent vs AI assistant: the difference and which to choose

An AI assistant answers questions in a chat: you ask, it writes an answer, and you do the rest. An AI agent does the whole job: it works with email and files, remembers people and agreements, acts on a schedule and messages you first. The same model does the thinking for both; what differs is what is built around it.

Read the article
  • What the difference is
  • A comparison table
  • One task, two approaches
  • What to choose
Neko in a bright workshop compares four robot agents on a table, a robot cat among them
Article · 7 min

Which AI agent is best: types of agents and how to choose yours

There is no single best AI agent for every case: agents come in four types, and each suits its own owner. They are cloud builders, the open-source agents OpenClaw and Hermes, your own agents on Claude Code or Codex, and ready-made turnkey agents. You choose by five questions: where the data lives, who fixes it, what it costs per month, whether it works in Telegram and which AI subscription it accepts.

Read the article
  • Four types of agents
  • Five questions to choose by
  • A comparison table
  • What to ask before buying
Neko in a bright office hands a stack of tasks to a robot cat sitting at its own desk with a laptop and a badge
Article · 7 min

AI employee: what it does, where a human is needed and what it costs

An AI employee is an AI agent given part of your regular work: correspondence, reminders, documents, summaries and reports. It works around the clock on a server, remembers people and agreements and runs tasks on a schedule. Money, promises to clients in your name and decisions with a high cost of error stay with a human. Per month an AI employee costs a server of about 1,500 RUB, an AI subscription and support.

Read the article
  • What the agent covers
  • Where a human is needed
  • Monthly cost
  • How to set tasks
In the evening Neko closes her laptop, while a robot cat with a phone sits on a running home server and keeps working
Article · 7 min

Local AI agent: on your own computer or on a server

A local AI agent is one that runs on your hardware rather than at a service. People mean two different things by it: a local model, when the neural network itself runs on your computer, and a local harness, when the agent program, files and memory live with you while a cloud model such as Claude or GPT does the thinking. An agent that works around the clock needs a computer that stays on, so it usually goes on your own server (VPS), about 1,500 RUB a month, rather than a laptop.

Read the article
  • Local model vs harness
  • An agent on a PC
  • Why you need a server
  • Where data is stored
In a small coffee shop Neko takes one sticky note with a task off a board and hands it to a robot cat holding a stopwatch
Article · 7 min

Implementing AI agents in a small business: a one-month plan and how to measure the gain

You can implement an AI agent in a small business within a month and without a separate project. Take one repeating routine, install the agent, spend two weeks fine-tuning it in ordinary chat and count how many hours a week it takes off your plate. No spec or integrations are needed at the start: an agent in Telegram only needs your explanations and access to the right files. Each month it costs a server of about 1,500 RUB and an AI subscription.

Read the article
  • The first routine
  • Two weeks of fine-tuning
  • How to measure the gain
  • Implementation mistakes
Neko types commands in a terminal in the evening, while a robot cat next to her studies an unfolded manual over an open box of parts, puzzled
Article · 7 min

Hermes Agent: what it is, how to install it and what you maintain yourself

Hermes Agent is a free, open-source AI agent from Nous Research under the MIT license. You install it on your computer or server, connect a model and a messenger, and it carries out tasks, keeps memory of past conversations and creates skills from solved tasks on its own. Installation is one terminal command or the Hermes Desktop app for Windows and macOS, followed by the “hermes setup” command. The software is free, while the model, the server and the upkeep stay with the owner.

Read the article
  • What Hermes can do
  • Installation step by step
  • Which models fit
  • What you maintain yourself
Neko works in AdsPower in the evening: the monitor shows a list of browser profiles, and the robot cat next to her points a paw at one of the rows
Article · 8 min

AdsPower: what this browser is and how to set up a profile and proxy

AdsPower is an antidetect browser: every profile has its own cookies, browser fingerprint and proxy, and websites see the profiles as different devices. That lets you run client accounts, team access and test accounts on one computer without mixing them up. AdsPower runs on Windows, macOS and Linux, two profiles are free forever, and you bring your own proxies: HTTP, HTTPS, SOCKS5 or SSH.

Read the article
  • What AdsPower is
  • A profile in five steps
  • Which proxies to pick
  • Common mistakes
Neko plugs a cable into a small home server in the evening, wires run from it to colourful chat bubbles, a robot cat sits on the server holding a padlock, and a red toy lobster stands nearby
Article · 8 min

OpenClaw: what it is, how to install it and connect Telegram

OpenClaw is a free, open-source AI assistant under the MIT license. It runs on your computer or server and answers in the messengers you already use: Telegram, WhatsApp, Discord, Slack, Signal and others. Installation is one terminal command, then a setup wizard connects the model, which takes about five minutes according to the docs. The software is free, while the model, the server, updates and security stay with the owner.

Read the article
  • What OpenClaw can do
  • Installation step by step
  • Telegram and a Claude subscription
  • What you maintain yourself
Neko on a sofa in the evening with her phone compares two helpers on the table: a plain grey bot waits with a single message bubble, while a robot cat juggles a calendar, an email, a document and a reminder
Article · 7 min

AI bot for Telegram or an AI agent: the difference and which to choose

An AI bot for Telegram is a bot that passes your message to a neural network and sends its answer back to the chat. An AI agent in Telegram looks the same but works differently: it lives on a server, remembers people and tasks, works with your files and email and messages you on a schedule. To ask, translate or draft a text, a bot is enough. To run your affairs every day, you need an agent.

Read the article
  • What an AI bot in Telegram is
  • How an agent differs from a bot
  • Comparison and prices
  • When a bot is enough
Neko at a workbench in the evening compares two glowing cores, orange and teal, while a robot cat with an open chest hatch waits next to her for one of them
Article · 7 min

Claude Code or Codex: which engine to choose for a personal AI agent

Claude Code by Anthropic and Codex by OpenAI are two harnesses that turn a model into an agent: they give it files, commands and tools on a server. For a personal agent the main choice between them is the subscription: Claude Code runs on a Claude Pro or Max plan, Codex on a ChatGPT plan, in a server terminal from Plus and up. The agent's memory, Telegram and schedule do not depend on the engine, and in NekoAgent you can switch the engine later right from the chat.

Read the article
  • What Claude Code and Codex are
  • Which subscription you need
  • How they differ day to day
  • Can you switch engines
Neko in a bright loft writes an instruction card and hands a finished glowing card to a robot cat holding an open card file with colored tabs
Article · 7 min

Claude Code skills: how to teach an AI agent a new task

A Claude Code skill is a folder with a SKILL.md file: a short description of when it is needed and step-by-step instructions. The agent sees the descriptions of all its skills and opens the right one by itself when a task matches the description, reading the full text only at that moment. This way you teach the agent a recurring task once, such as replying to a review or a weekly report, and never explain it again. Codex by OpenAI understands the same format.

Read the article
  • What a skill is
  • How the agent picks a skill
  • Business examples
  • Skills and plugins
Neko relaxes in an armchair with a tablet while a robot cat sorts letters into trays on its own and shows her a card with a question mark and a green check; a low glass barrier glows next to the desk
Article · 7 min

Autonomous AI agents: what an agent does on its own and when it asks a human

An autonomous AI agent performs tasks without a command for every step: on a schedule and by the rules you gave it. It sends summaries and reminders, prepares drafts, records decisions in its memory and messages you first. Money, messages to people on your behalf and deleting data are best kept subject to your confirmation, even for an autonomous agent. Limits are set in two ways: with rules in your chat with the agent and with the access you give it.

Read the article
  • What the agent does alone
  • When it asks a human
  • How to set limits
  • Several agents
Neko in her kitchen in the evening looks at a chat message on her phone; a robot cat on the counter holds a glowing card with a bell and a checklist; on the fridge are a calendar with circled dates, a child's drawing and a shopping list, next to a grocery bag and a folder of documents
Article · 7 min

Personal AI assistant: errands, family and reminders in one chat

A personal AI assistant is an agent that runs your errands in Telegram: it remembers agreements, dates and documents, reminds you about them on its own and takes tasks by text or voice. It differs from an AI chat on your phone in two ways: it has a lasting memory about you and it messages you first when a deadline approaches.

Read the article
  • What to give the assistant
  • How it differs from a chat
  • An assistant for the family
  • Personal data